Lightening Network for Low-Light Image Enhancement Paper Reading Notes

• This is an article on supervised dark image enhancement in the 2022 TIP journal

• The network structure is shown in the figure:
  

• The network structure of LBP is as follows:
  
  it is a bit convoluted, and the theory it is based on is as follows. That is to say, ordinary dark image enhancement is just to follow the L1 red arrow in the figure below and estimate a bright image from the dark image. But in fact, there is still some gap between this bright picture and the real bright picture. How to make up for it? You can further learn an $R_{LL}$to $R_{NL}$Mapping, thereby adding this predicted residual to the initially predicted bright image, can make the final enhancement result closer to the bright image.
  
  Here $L_1$, $L_2$, and $D$ as shown below
  

• The FA module is shown in the figure below, which is a channel attention module:
  

• The loss function is the sum of TV loss and SSIM loss of the enhanced result

• The experimental results on the LOL data set look pretty good:
  
  

• The inspiration is that the role of back projection and residual learning is quite large. At the same time, the network is iteratively enhanced, which means that there are many 3-channel enhancement results in the middle. This is compared to other networks where only features are in the middle. More explainable.